We walk through a plan-first workflow that asks Copilot to survey the codebase, propose a step-by-step solution, and only then switch to agent mode to implement. That simple change lifts code quality, reduces rework, and keeps diffs aligned with architecture. We also dig into custom agents with scoped permissions and roles: a testing agent for unit coverage, a docs agent for READMEs, a refactor agent limited to certain directories. Prompt craft matters, but stable configuration, coding standards, and CI guardrails matter more. Think of it as turning best practices into reusable instructions that AI can follow every time.

Johan shares field notes from consulting and charity projects: where AI saves hours, where it still stumbles, and how code generation quality has improved over the past year. We explore SpecKit’s promise for scaffolding and simpler apps, acknowledge its preview status and quota costs, and outline how to adopt incrementally in legacy systems. Along the way, we cover open source tracking via changelogs and issues, integrating agents into CI/CD, and designing a workflow that is auditable, secure, and team-friendly.

Text Us About the Show